Search

This is a response to Nancy Cook's Readers' Reaction to "Measures to Summarize and Compare the Predictive Capacity of Markers."

Previous studies of risk factors for acute graft-versus-host disease (GVHD) involved patients receiving predominantly single-agent prophylaxis. Therefore, a retrospective analysis was performed on 446 patients, from a single institution, who received transplants of marrow from HLA-identical siblings and the combination of cyclosporine (CSP) and methotrexate (MTX) to determine risk factors for acute GVHD associated with this more effective form of GVHD prophylaxis. The incidences of Grades II-IV and Grades III-IV (severe) acute GVHD were 35% and 16%, respectively. Increased clinical grades of acute GVHD in patients without advanced malignant disease were associated with a decreased survival. In a multivariate Cox regression analysis, risk factors associated with the onset of Grades II-IV acute GVHD were sex mismatch and donor parity (P = .001), increased dose of total body irradiation (TBI) (P = .001), and reduction to less than 80% of the scheduled dose of MTX (P = .02) or CSP (P = .02). The multivariate analysis indicated a relative risk of 1.37 for acute GVHD in a group defined as having advanced malignant disease at transplant; however, this difference failed to reach conventional levels of statistical significance (P = .07). Reduction of MTX and CSP occurred in up to 36% and 44% of patients, respectively, primarily because of renal or hepatic dysfunction. The periods of increased risk for the onset of acute GVHD were up to 1 week after a reduction of MTX and 2 weeks after a reduction in CSP. When only patients who developed Grades II-IV acute GVHD were considered, the more severe acute GVHD of Grades III-IV was associated with increased patient age of 40 years or greater (P = .05) and dose reductions of CSP (P = .008). Serologic status of patient and donor for cytomegalovirus (CMV), HLA antigens in the A and B loci, and isolation in a laminar air flow room during marrow transplantation, all previously identified as risk factors for acute GVHD, were not confirmed as risk factors in this study population. The toxicity of MTX and CSP and the development of acute GVHD from inadequate immunosuppression because of dose reduction warrants further trials with potentially less toxic immunosuppressive agents. Risk factors for acute GVHD should be considered in clinical management and in the design of clinical trials.

Sixty-eight patients with moderate (n = 15) or severe (n = 53) aplastic anemia were entered into a prospective, randomized, two-arm treatment study comparing antihuman thymocyte globulin (ATG), lower-dose methylprednisolone (LDM) and oxymetholone to ATG, higher-dose methylprednisolone (HDM), and oxymetholone. There were no differences between the two groups when comparing age, sex, etiology of aplasia, disease duration, severity of aplasia, or pretherapy granulocyte counts. Side effects of LDM and HDM were similar. Of the 64 patients evaluable for response to therapy, 12 of 33 (36%) who received LDM had complete, partial, or minimal responses compared with 15 of 31 patients (48%) who received HDM (P = .33). Actuarial survival at 4 years is 43% for patients in the LDM group and 47% for patients in the HDM group (P = .99). Causes of death included hemorrhage, infection, evolution to acute leukemia, and complications of subsequent bone marrow transplantation. Long-term complications included paroxysmal nocturnal hemoglobinuria (n = 3), evolution to myelodysplasia or acute leukemia (n = 6), and recurrent aplasia (n = 6). We were unable to show a significant difference in toxicity, response rate, or survival for patients treated with ATG, oxymetholone, and LDM compared with patients who received ATG, oxymetholone, and HDM.

The efficiency of extracting hexavalent chromium from a silica matrix with a 3% sodium carbonate‐1% sodium hydroxide solution was demonstrated using an “in‐house”; reference material, and this extraction procedure was successfully applied to the determination of hexavalent chromium in some contaminated soils from Hudson County, New Jersey.

Summary. Nineteen patients with leukaemia, preleukaemia, and aplastic anaemia were treated by marrow transplantation from HLA-identical siblings. All were given postgrafting immunosuppression with a combination of methotrexate (days 1,3,6 and 11) and cyclosporin (days -1 to 180). In an attempt at reducing cyclosporin-associated toxicity, we explored whether the cyclosporin dose during the first 2 weeks could be decreased by 50% (from 3.0 to 1.5 mg/kg/d intravenously) without adversely affecting the incidence, onset, and severity of acute graft-versus-host disease (GVHD) and overall survival. Results from this pilot study were compared to those of a matched cohort of 38 patients given a standard dose of 3.0 mg cyclosporin/kg/d starting on day -1. The cumulative incidence of grade II and III acute GVHD in the ‘low dose’ cyclosporin group was 42% compared to 51% in the ‘standard dose’ group (P= 0.60). Three-year survival was 63% and 54% respectively (P=0.59). Patients receiving the reduced cyclosporin dose during the first 14 d appeared to have less hepatotoxicity, and the methotrexate and cyclosporin doses administered were closer to the doses intended per protocol. We suggest that‘low dose’cyclosporin from day -1 to day 15 postgrafting might be as effective as ‘standard dose’ cyclosporin during that time period for the prevention of acute GVHD.

Critical current measurements up to 8 T, using the facility MARISA, on cables for LHC (Large Hadron Collider) dipole magnets are reported. The samples were fed using an inductive method. The features of the method are discussed. AC measurements on the strands were performed in order to detect the proximity effect between the filaments. >

SUMMARY In the context of estimating ,3 from the regression model P ( YI X), relating response Y to covariates X, suppose that only a surrogate response S is available for most study subjects. Suppose that for a random subsample of the study cohort, termed the validation sample, the true outcome Y is available in addition to S. We consider maximum likelihood estimation of ,3 from such data and show that it is nonrobust to misspecification of the distribution relating the surrogate to the true outcome, P(S I Y, X). An alternative semiparametric method is also considered, which is nonparametric with respect to P(S I Y, X). Large-sample distribution theory for maximum estimated likelihood estimates is developed. An illustrative example is presented.

Biomarkers that can be used in combination with established screening tests to reduce false positive rates are in considerable demand. In this article, we present methods for evaluating the diagnostic performance of combination tests that require positivity on a biomarker test in addition to a standard screening test. These methods rely on relative true and false positive rates to measure the loss in sensitivity and gain in specificity associated with the combination relative to the standard test. Inference about the relative rates follows from noting their interpretation as conditional probabilities. These methods are extended to evaluate combinations with continuous biomarker tests by introducing a new statistical entity, the relative receiver operating characteristic (rROC) curve. The rROC curve plots the relative true positive rate versus the relative false positive rate as the biomarker threshold for positivity varies. Inference can be made by applying existing ROC methodology. We illustrate the methods with two examples: a breast cancer biomarker study proposed by the Early Detection Research Network (EDRN) and a prostate cancer case-control study examining the ability of free prostate-specific antigen (PSA) to improve the specificity of the standard PSA test.

The predictive capacity of a marker in a population can be described using the population distribution of risk (Huang et al. 2007; Pepe et al. 2008a; Stern 2008). Virtually all standard statistical summaries of predictability and discrimination can be derived from it (Gail and Pfeiffer 2005). The goal of this paper is to develop methods for making inference about risk prediction markers using summary measures derived from the risk distribution. We describe some new clinically motivated summary measures and give new interpretations to some existing statistical measures. Methods for estimating these summary measures are described along with distribution theory that facilitates construction of confidence intervals from data. We show how markers and, more generally, how risk prediction models, can be compared using clinically relevant measures of predictability. The methods are illustrated by application to markers of lung function and nutritional status for predicting subsequent onset of major pulmonary infection in children suffering from cystic fibrosis. Simulation studies show that methods for inference are valid for use in practice.

Kaplan–Meier and cumulative incidence functions are not sufficient descriptive devices for studies that have multiple time-to-event endpoints. For example, in cancer treatment research the probability of tumor recurrence conditional on not having died from treatment-related toxicities and the prevalence of graft-versus-host disease among leukemia-free patients surviving a bone marrow transplant are of interest. These quantities can be estimated nonparametrically using simple functions of several Kaplan–Meier and cumulative incidence estimates for events with possibly dependent risks. We derive asymptotic distribution theory for such functions by representing Kaplan–Meier, cumulative incidence, and cumulative hazard estimators as sums of iid random variables. Variance estimation also follows directly from this representation. Two-sample test statistics with asymptotic null distribution theory are presented. Several examples illustrate the utility of these results.

The activation signals leading to proliferation of normal and leukemic CD3+ large granular lymphocytes (LGL) were studied in vitro. Anti-CD3 monoclonal antibody (MoAb) alone (P less than .01) and recombinant interleukin-2 (IL-2) alone (P less than .01) caused significant stimulation of peripheral blood mononuclear cells (PBMC) from four CD3+ LGL leukemia patients, as measured in a 3H-thymidine incorporation assay. Recombinant interleukin-4 (IL-4) alone had no effect (P = .11). The combination signals of anti-CD3 MoAb and either IL-2 or IL-4 produced a proliferative response greater than anti-CD3 MoAb alone (P less than .01) or lymphokine alone (P less than .01). Leukemic LGL, purified by two-color sorting, were subsequently activated by anti-CD3 MoAb and IL-2 and assessed for DNA content by viable Hoechst No. 33342 (HO) staining. Results of these studies demonstrated that leukemic LGL were stimulated directly by anti-CD3 MoAb and IL-2, with the percentage of cells in cell cycle (S + G2/M) ranging from 16% to 72%. Normal CD3+ LGL were also stimulated to enter the cell cycle by anti-CD3 and IL-2. These results show that leukemic LGL proliferate in vitro after activation through the T-cell receptor and/or lymphokine.

Mismeasurement of covariate data is a frequent problem in statistical data analysis. However, when true and mismeasured data are obtained for a subsample of the observations, it is possible to estimate the parameters relating the outcome to the covariate of interest. Maximum likelihood methods that rely on parametric models for the mismeasurement have not met with much success. Realistic models for the mismeasurement process are difficult to construct; the form of the likelihood is often intractable and, more important, such methods are not robust to model misspecification. We propose an easily implemented method that is nonparametric with respect to the mismeasurement process and that is applicable when mismeasurement is due to the problem of missing data, errors in variables, or use of imperfect surrogate covariates. Specifically, denote the outcome variable by Y, the covariate data subject to mismeasurement by X, and the remaining covariates, including perhaps surrogates or mismeasured values ...

We propose an estimator of the prevalence of a transient condition among surviving patients using right censored data. The prevalence of opportunistic infection among surviving AIDS patients and the probability of being in tumour response following cancer therapy conditional on being alive are two examples of such functions. In essence these functions describe a major aspect of the quality of life for surviving patients and may be useful when viewed in conjunction with the survival curves themselves. The method is illustrated using data from a randomized trial of bone marrow transplant patients where the prevalence of chronic graft-versus-host disease is of interest. The non-parametric estimator which we have proposed is contrasted with estimators derived from Markov and semi-Markov models.

Less than 40% of the patients who could benefit from marrow transplantation have an HLA-matched relative who can serve as a donor. For this reason, several centers have explored marrow transplantation from other categories of donors. This retrospective study analyzes the results of marrow transplantation for 52 patients receiving grafts from HLA-A,B,DR,Dw-phenotypically matched, MLC-compatible, unrelated volunteer donors compared to a disease, disease-stage, and age-matched cohort of 104 patients transplanted from HLA-genotypically identical sibling donors. The patients transplanted from unrelated donors had an increased incidence of grade II-IV acute graft-versus-host disease compared to patients transplanted from related donors (79% vs. 36%, P much less than 0.001). However, the probability of relapse-free survival appears similar in the two groups (P = 0.39 over all, with estimates of 41% vs. 46% at 1 year). We conclude from this preliminary data that marrow transplantation from HLA-matched unrelated donors should be considered in most, if not all, circumstances where transplantation from an HLA-matched sibling would be indicated if such a donor were available.

SUMMARY We propose a cumulative weighted difference in the Kaplan-Meier estimates as a test statistic for equality of distributions in the two-sample censored data survival analysis problem. For stability of such a statistic, the absolute value of the possibly random weight function must be bounded above by a multiple of (C-)1/2+6 where 1 - C - is the left continuous censoring distribution function and 6 > 0. For these weighted Kaplan-Meier (WKM) statistics, asymptotic distribution theory is presented along with expressions for the efficacy under a sequence of local alternatives. A simple censored data generalization of the two-sample difference in means test (z-test) is a member of this class and in large samples is seen to be quite efficient relative to the popular log-rank test under a range of alternatives including the proportional hazards alternative. Optimal weight functions are also calculated. The optimal WKM statistic is as efficient as the optimal weighted log-rank statistic for any particular sequence of local alternatives. Stratified statistics and trend statistics are also presented.

The recent epidemiologic and clinical literature is filled with studies evaluating statistical models for predicting disease or some other adverse event. Risk stratification tables are a new way to evaluate the benefit of adding a new risk marker to a risk prediction model that includes an established set of markers. This approach involves cross-tabulating risk predictions from models with and without the new marker. In this article, the authors use examples to show how risk stratification tables can be used to compare 3 important measures of model performance between the models with and those without the new marker: the extent to which the risks calculated from the models reflect the actual fraction of persons in the population with events (calibration); the proportions in which the population is stratified into clinically relevant risk categories (stratification capacity); and the extent to which participants with events are assigned to high-risk categories and those without events are assigned to low-risk categories (classification accuracy). They detail common misinterpretations and misuses of the risk stratification method and conclude that the information that can be extracted from risk stratification tables is an enormous improvement over commonly reported measures of risk prediction model performance (for example, c-statistics and Hosmer-Lemeshow tests) because it describes the value of the models for guiding medical decisions.

The classification accuracy of a continuous marker is typically evaluated with the receiver operating characteristic (ROC) curve. In this paper, we study an alternative conceptual framework, the "percentile value." In this framework, the controls only provide a reference distribution to standardize the marker. The analysis proceeds by analyzing the standardized marker in cases. The approach is shown to be equivalent to ROC analysis. Advantages are that it provides a framework familiar to a broad spectrum of biostatisticians and it opens up avenues for new statistical techniques in biomarker evaluation. We develop several new procedures based on this framework for comparing biomarkers and biomarker performance in different populations. We develop methods that adjust such comparisons for covariates. The methods are illustrated on data from 2 cancer biomarker studies.

Predicting risk of cancer for individuals has long been a goal of medical research. If an individual’s risk could be predicted, then prevention and screening modalities could be targeted toward those at meaningfully high risk. This approach is not only more cost efficient than targeting the whole population but also more ethical, at least when interventions are burdensome to the individual. The quest for risk predictors has been revitalized with the emergence of technologies that measure genetic information and other molecular and physiological attributes of the individual. In this issue of the Journal, Gail (1) asks to what extent newly discovered associations between seven single-nucleotide polymorphisms (SNPs) and incidence of breast cancer can improve assessment of breast cancer risk. Comparisons are made with models that employ standard clinical factors to evaluate the incremental value of the SNPs for prediction over the standard clinical information. Using estimated relative risks and allele frequencies, Gail finds that the SNPs are expected to have a small effect on the capacity of prediction models to distinguish women who will and will not develop breast cancer. Because he assumes best-case scenarios, his results probably provide upper limits on expected increments in risk prediction with SNPs. He postulates that many more SNPs with these levels of association with breast cancer will need to be discovered to substantially improve risk prediction. Gail’s arguments demonstrate that the sample sizes needed to discover an adequate number of SNPs will need to be very large indeed. Although his calculations are based on many assumptions, they provide a good place from which to start the discussion about what types of markers and studies will be needed to make progress in this field. Gail uses the area under the receiver operating characteristic (ROC) curve (AUC) to summarize and compare prediction models. Although this is a popular statistical approach with a long history (2), there has been considerable criticism of it, with recent criticisms coming from the cardiovascular research community (3, 4). Reexamining the role of AUC has been motivated in part by frustration at not being able to identify valuable biomarkers on the basis of AUC. The AUC is often interpreted as the probability of correct ordering—correct in the sense that when comparing the risk predictions for two subjects, only one of whom develops breast cancer, the risk calculated for the breast cancer subject is the larger of the two values. The first criticism of the AUC is that this probability is not a clinically relevant way of summarizing predictive performance. Subjects do not present to the clinic in pairs, and the problem is not to determine which of the pair will develop cancer. A second criticism is that the scale is somewhat deceptive. A substantial gain in performance may not result in a substantial increase in AUC. For example, suppose that the sensitivity changes from 40% to 90% but only over the range of specificities corresponding to 90% – 100%. This is an enormous improvement in performance: while maintaining specificity at 100%, now 90% rather than 40% of cancers can be predicted. However, the change in AUC is only 0.05. Although an extreme example, it illustrates the point. These two criticisms of AUC apply generally, not solely to risk prediction. The AUC really is a poor metric for evaluating markers for disease diagnosis, screening, or prognosis. The third criticism, which is specific to risk prediction, is that the AUC, and indeed the ROC curve itself, hides the values of risk calculated for subjects in the population. Indeed, the risk values are not visible from the ROC curve or the related curves in figure 2 of Gail (1). Moreover, the same ROC curve results if risk values are transformed monotonically, say, multiplied by a factor of 10, yet the clinical implications of these risk values would be very different. The key question in evaluating risk prediction models concerns the number of subjects who are identified as being at high risk. Does a model that includes SNPs identify a substantially larger number of women at high risk for breast cancer who might therefore benefit from an intervention? In other words, does it do a better job than a model without SNPs at risk stratifying the population? The population distributions of absolute risk derived from the two models can be displayed to allow this sort of assessment. At any chosen high-risk threshold, the proportions of subjects with risks above that threshold can be compared. We refer to these plots as “predictiveness curves” (5, 6). Cook (3) tabulates these proportions for specific thresholds considered therapeutically relevant in preventing cardiovascular events. Unfortunately, it is not possible to determine the population distributions of absolute risk for the models described in Gail (1). The distributions of relative risk are shown instead. Gail notes that to derive the absolute risk distribution from the relative risk distribution, one needs to know the absolute risk in the baseline group, those with the lowest level of risk for all factors in the model, denoted by Gail with the letter k. The effect of k would be to shift the relative risk distribution shown in his figure 1 by log(k) to arrive at the distribution curve for absolute risk. Because the models differ in risk factors included, the baseline group varies across models and so too does the corresponding risk, k. This means that the curves in figure 1 would need to be shifted by different degrees to assemble the absolute risk distributions from them. In conclusion, the comparison of relative risk distributions does not give direct information about the comparison of absolute risk distributions, which is of key interest for comparing risk prediction models. Interestingly, the absolute risk distributions could have been calculated by Gail if population incidence rates were specified. In particular, the absolute risk in the baseline group, k, for each model is simply the population incidence divided by the average relative risk. Because k is the factor that links the relative risk distribution to the absolute risk distribution, values for age-specific incidence of breast cancer could therefore be used in conjunction with Gail’s calculations to determine the age-specific risk distributions for women using models with and without SNPs. The age-specific proportions of women identified at high risk could then be compared across models. This would be an interesting exercise to complement Gail’s calculations. Appropriate evaluation of risk prediction models requires specification of a risk threshold for defining individuals as high risk. What high-risk threshold should be used in the breast cancer setting? A consensus on this fundamental question does not exist at present. The choice depends on costs and benefits associated with interventions that will be employed for women designated as high risk. Tamoxifen therapy and screening with magnetic resonance imaging are among the set of options for breast cancer. Medical decision theory provides an explicit solution for high-risk designation in terms of 1) the net benefit, B, of being classified as high risk if, in the absence of intervention, one is destined to develop breast cancer, and 2) the net cost, C, of being classified as high risk if, in the absence of intervention, one is destined not to develop breast cancer. The risk threshold at which expected benefit exceeds expected cost is C/(C + B) (7). The higher the cost:benefit ratio, the higher the optimal threshold. Cardiovascular consensus groups (8) have determined risk thresholds based on costs and benefits of different therapy options. Corresponding guidelines for defining high risk in the context of breast cancer prevention must be developed. We need them to gauge the value of risk prediction models. Risk thresholds might be chosen to vary with factors such as age, recognizing that costs and benefits of high-risk interventions are not uniform across the population. Moreover, in practice each woman may have her own tolerance for risk that could vary from guidelines developed by consensus groups. Gail’s ROC analysis indicates that even under optimistic assumptions, SNPs—or, for that matter, other risk factors with moderate relative risks—are unlikely to substantially improve current algorithms for breast cancer risk prediction. The same conclusion may well hold with analyses that focus on proportions of high-risk (or low-risk) women identified. Indeed, this has been the experience in cardiovascular research. Biomarkers such as C-reactive protein (CRP) and high-density lipoproteins that do not increase AUC statistics do not appear to improve risk stratification either, at least when considering the population as a whole (9). Subsets of the population may, however, benefit from information in these markers. Ridker and Cook (10) report that for subjects at intermediate risk according to standard risk factors, CRP can further stratify a large fraction of subjects into high- and low-risk categories. Similarly, for breast cancer risk prediction SNPs and biomarkers may have their greatest impact on subpopulations. Risk stratification is not the only component of prediction model evaluation. As Gail notes, calibration is of paramount importance. A well-calibrated model ensures that the calculated risks reflect the actual proportions of subjects who develop disease. Also of interest is the accuracy of risk classifications, defined as the proportion of women who develop breast cancer who are classified as high risk and the proportion of women who do not develop breast cancer who are classified as low risk (5). These can also be calculated using the absolute risk distribution, a fact previously noted by Gail and Pfeiffer (11).

The concept of covariate adjustment is well established in therapeutic and etiologic studies. However, it has received little attention in the growing area of medical research devoted to the development of markers for disease diagnosis, screening, or prognosis, where classification accuracy, rather than association, is of primary interest. In this paper, the authors demonstrate the need for covariate adjustment in studies of classification accuracy, discuss methods for adjusting for covariates, and distinguish covariate adjustment from several other related, but fundamentally different, uses for covariates. They draw analogies and contrasts throughout with studies of association.

This paper presents a comprehensive overview of the surface deformation retrieval capability of the Differential Synthetic Aperture Radar Interferometry (DInSAR) algorithm, referred to as Small BAseline Subset (SBAS) technique, in the context of active volcanic areas. In particular, after a brief description of the algorithm some experiments relevant to three selected case-study areas are presented. First, we concentrate on the application of the SBAS algorithm to a single-orbit scenario, thus considering a set of SAR data composed by images acquired on descending orbits by the European Remote Sensing (ERS) radar sensors and relevant to the Long Valley caldera (eastern California) area. Subsequently, we address the capability of the SBAS technique in a multipleorbit context by referring to Mt. Etna volcano (southern Italy) test site, with respect to which two different ERS data set, composed by images acquired both on ascending and descending orbits, are available. Finally, we take advantage of the capability of the algorithm to work in a multi-platform scenario by jointly exploiting two different sets of SAR images collected by the ERS and the Environment Satellite (ENVISAT) radar sensors in the Campi Flegrei caldera (southern Italy) area. The presented results demonstrate the effectiveness of the algorithm to investigate the deformation field in active volcanic areas and the potential of the DInSAR methodologies within routine surveillance scenario.

One hundred forty-seven consecutive patients with leukemia, myelodysplastic syndrome, or aplastic anemia were treated by marrow grafts from genotypically HLA-identical siblings (n = 122) or HLA- haploidentical family members (n = 25). Haploidentical recipients differed from their donors for no more than one HLA locus on the nonshared haplotype. All were given postgrafting immunosuppression with a combination of methotrexate and cyclosporine. In a randomized study we explored whether prednisone administered from day 0 through 35 along with methotrexate/cyclosporine could improve prevention of acute graft- versus-host disease (GVHD). The GVHD incidence in patients not given prednisone was comparable with that previously reported with methotrexate/cyclosporine. Unexpectedly, significant increases in acute and also chronic GVHD were seen in HLA-identical recipients administered prednisone, but not in the small number of patients administered HLA-nonidentical grafts. However, the resultant increase in transplant-related mortality in patients administered prednisone was offset by an increase in leukemic relapse in patients not administered prednisone, presumably related to the absence of a graft-versus- leukemia effect. Therefore, overall disease-free survival of the two groups of patients was comparable, with slightly more than 50% of the patients being alive at more than 2 years after transplantation. We speculated that prednisone adversely affected GVHD prophylaxis, interfering with methotrexate's cell cycle-dependent suppression of donor lymphocyte proliferation in response to host antigens. In a pilot study we explored whether beginning prednisone on day 15, after completion of methotrexate administration, would avoid this adverse effect. The GVHD incidence in patients administered methotrexate/cyclosporine along with “late” prednisone was comparable with that in patients not administered prednisone. We conclude that methotrexate/cyclosporine is effective in decreasing the incidence of grade II through IV GVHD, and that the addition of prednisone to this regimen is not beneficial in recipients of HLA-identical marrow grafts.

An ELISA was used to study long-term immunity and immunization responses to tetanus toxoid in 48 bone marrow transplant recipients. Among patients who were seropositive to tetanus before transplant, 51% had lost their seropositivity 1 year later. All patients who were not reimmunized with tetanus toxoid were seronegative 2 years after transplant. All patients who were seronegative before transplant remained seronegative 1 year later regardless of the donor's serologic status. There was no difference in the ability to remain seropositive to tetanus toxoid between patients with and without chronic graft-versus-host disease. Of 21 patients immunized with one dose of tetanus toxoid 1 year after transplant, 14 were seronegative at the time of immunization (response rate, 64%). At 1 year after immunization, 7 remained seropositive. Ten patients were reimmunized with two doses of tetanus toxoid. All responded and 90% remained seropositive 1 year later. When 21 patients were primarily immunized with three doses of tetanus toxoid, all patients seronegative at immunization responded and all tested patients remained seropositive 2 years later. The immunization responses were significantly superior in patients receiving three doses compared with those who received one. Reimmunization with tetanus toxoid of long-term survivors after marrow transplant seems necessary. A three-dose immunization schedule is recommended to obtain an adequate immune response.

To the Editor: Current statistical approaches for evaluation of risk prediction markers are unsatisfactory. We applaud Cook’s criticisms of the c-index, or area under the receiver operating characteristic curve. This index is based on the notion of pairing subjects, one with poor outcome (eg, cardiovascular event within 10 years) and one without, and determination of whether the risk for the former (ie, the case) is larger than the risk for the latter (ie, the control). This probability of correct ordering of risks is not a relevant measure of …

In case-control studies evaluating the classification accuracy of a marker, controls are often matched to cases with respect to factors associated with the marker and disease status. In contrast with matching in epidemiologic etiology studies, matching in the classification setting has not been rigorously studied. In this article, we consider the implications of matching in terms of the choice of statistical analysis, efficiency, and assessment of the incremental value of the marker over the matching covariates. We find that adjustment for the matching covariates is essential, as unadjusted summaries of classification accuracy can be biased. In many settings, matching is the most efficient covariate-dependent sampling scheme, and we provide an expression for the optimal matching ratio. However, we also show that matching greatly complicates estimation of the incremental value of the marker. We recommend that matching be carefully considered in the context of these findings.

Diagnostic tests, medical tests, screening tests, biomarkers, and prediction rules are all types of classifiers. This chapter introduces methods for classifier development and evaluation. We first introduce measures of classification performance including sensitivity, specificity, and receiver operating characteristic (ROC) curves. We then review some issues in the design of studies to assess and compare the performance of classifiers. Approaches for using the data to estimate and compare classifier accuracy are then introduced. Next, methods for combining multiple classifiers into a single classifier are presented. Lastly, we discuss other important aspects of classifier development and evaluation. The methods presented are illustrated with real data.

We have investigated the ongoing deformation at the Campi Flegrei caldera (Italy), from 2002 to the end of 2006, by analyzing ENVISAT ASAR IS-2 data. The study has been performed by exploiting the SBASDInSAR algorithm that allows us to investigate the temporal evolution of deformation at both low and full spatial resolution scales. The low resolution analysis highlighted the renewed volcanic activity that started on mid-2005; moreover, the obtained results have been confirmed by the leveling data collected by the Osservatorio Vesuviano. The full resolution analysis allowed us to identify isolated subsiding structures, whose deforming behaviors were not detected at the low resolution spatial scale.

Developing biomarkers that can predict whether patients are likely to benefit from an intervention is a pressing objective in many areas of medicine. Recent guidance documents have recommended that the accuracy of predictive biomarkers, ie, sensitivity, specificity, and positive and negative predictive values, should be assessed. We clarify the meanings of these entities for predictive markers and demonstrate that generally they cannot be estimated from data without making strong untestable assumptions. Language suggesting that predictive biomarkers can identify patients who benefit from an intervention is also widespread. We show that in general one cannot estimate the chance that a patient will benefit from treatment. We recommend instead that predictive biomarkers be evaluated with respect to their ability to predict clinical outcomes among patients treated and among patients receiving standard of care, and the population impact of treatment rules based on those predictions. Ideally these entities are estimated from a randomized trial comparing the experimental intervention with standard of care.

Background: Triple negative breast cancers (TNBC) comprise 15-20% of all breast cancers and frequently present as interval cancers with high proliferative rates and increased risk of mortality. There is a clinical need for biomarkers for the early detection of TNBC to complement radiologic imaging. No plasma biomarkers for TNBC currently exist. The purpose of this study is to evaluate a panel of novel plasma biomarkers for TNBC for the discrimination of TNBC and benign breast disease as a crucial step in identifying a panel of plasma biomarkers for early detection. Methods: In a multicenter collaboration between the NCI EDRN and the CPTAC consortium, we conducted a prospective blinded phase II biomarker study that evaluated 76 candidate TNBC plasma biomarkers. Plasma samples collected at the time of diagnosis from 65 TNBC cases and 195 matched controls with benign breast disease without atypia were identified from multiple clinical sites. The samples were distributed as blinded aliquots to the biomarker laboratories for protein and autoantibody detection. Candidate protein (n=54) and autoantibody (n=22) biomarkers were selected and ranked prior to evaluation. All results were centrally analyzed. The sensitivity at 95% specificity was calculated for each biomarker. Effects of age, race, and specimen source on biomarkers were evaluated. Logistic regression was used to assess complementarity of biomarkers. The top three biomarkers underwent verification with an independent set of 60 TNBC cases and 180 matched controls from women undergoing mammography. Results: Statistically significant differences in case versus control signals were observed for 3 biomarkers with sensitivities of 17-23% at 95% specificity (p Conclusion: We have developed a pipeline strategy for the validation of plasma biomarkers for detection of breast cancer. At least three biomarkers for TNBC were confirmed in this study. Further evaluation of these biomarkers for early detection is ongoing. Citation Format: Karen S Anderson, Margaret Pepe, Jeffrey Marks, Paul Engstrom, Christos Patriotis, Richard Zangar, Steven Skates, Paul Lampe, Joshua LaBaer, Christopher I Li. A blinded multicenter phase II study of a panel of plasma biomarkers for the detection of triple negative breast cancer [abstract]. In: Proceedings of the Thirty-Seventh Annual CTRC-AACR San Antonio Breast Cancer Symposium: 2014 Dec 9-13; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2015;75(9 Suppl):Abstract nr P4-02-05.

No single biomarker for cancer is considered adequately sensitive and specific for cancer screening. It is expected that the results of multiple markers will need to be combined in order to yield adequately accurate classification. Typically, the objective function that is optimized for combining markers is the likelihood function. In this article, we consider an alternative objective function-the area under the empirical receiver operating characteristic curve (AUC). We note that it yields consistent estimates of parameters in a generalized linear model for the risk score but does not require specifying the link function. Like logistic regression, it yields consistent estimation with case-control or cohort data. Simulation studies suggest that AUC-based classification scores have performance comparable with logistic likelihood-based scores when the logistic regression model holds. Analysis of data from a proteomics biomarker study shows that performance can be far superior to logistic regression derived scores when the logistic regression model does not hold. Model fitting by maximizing the AUC rather than the likelihood should be considered when the goal is to derive a marker combination score for classification or prediction.

In this issue, Leening and colleagues discuss limitations and controversies surrounding the NRI and propose a systematic approach for presenting NRI analysis. The editorialists recommend that resea...

Non-parametric procedures such as the Wilcoxon rank-sum test, or equivalently the Mann-Whitney test, are often used to analyse data from clinical trials. These procedures enable testing for treatment effect, but traditionally do not account for covariates. We adapt recently developed methods for receiver operating characteristic (ROC) curve regression analysis to extend the Mann-Whitney test to accommodate covariate adjustment and evaluation of effect modification. Our approach naturally extends use of the Mann-Whitney statistic in a fashion that is analogous to how linear models extend the t-test. We illustrate the methodology with data from clinical trials of a therapy for Cystic Fibrosis.

Background: Markers that purport to distinguish subjects with a condition from those without a condition must be evaluated rigorously for their classification accuracy. A single approach for statistical evaluation and comparison of markers is not yet established. Methods: We suggest a standardization that uses the marker distribution in unaffected subjects as a reference. For an affected subject with marker value Y, the standardized placement value is the proportion of unaffected subjects with marker values that exceed Y. Results: We applied the standardization to 2 illustrative datasets. As a marker for pancreatic cancer, the CA-19-9 marker had smaller placement values than the CA-125 marker, indicating that CA-19-9 was the better marker. For detecting hearing impairment, the placement values for the test output (the marker) were smaller when the input sound stimulus was of lower intensity, which indicates that the test better distinguishes hearing-impaired from unimpaired ears when a lower intensity sound stimulus is used. Explicit connections are drawn between the distribution of standardized marker values and the receiver operating characteristic curve, one established statistical technique for evaluating classifiers. Conclusion: The standardization is an intuitive procedure for evaluating markers. It facilitates direct and meaningful comparisons between markers. It also provides a new view of receiver operating characteristic analysis that may render it more accessible to those as yet unfamiliar with it. The general approach provides a statistical tool to address important questions that are typically not addressed in current marker research, such as quantifying and controlling for covariate effects.

The statistical literature on assessing the accuracy of risk factors or disease markers as diagnostic tests deals almost exclusively with settings where the test, Y, is measured concurrently with disease status D. In practice, however, disease status may vary over time and there is often a time lag between when the marker is measured and the occurrence of disease. One example concerns the Framingham risk score (FR-score) as a marker for the future risk of cardiovascular events, events that occur after the score is ascertained. To evaluate such a marker, one needs to take the time lag into account since the predictive accuracy may be higher when the marker is measured closer to the time of disease occurrence. We therefore consider inference for sensitivity and specificity functions that are defined as functions of time. Semiparametric regression models are proposed. Data from a cohort study are used to estimate model parameters. One issue that arises in practice is that event times may be censored. In this research, we extend in several respects the work by Leisenring et al. (1997) that dealt only with parametric models for binary tests and uncensored data. We propose semiparametric models that accommodate continuous tests and censoring. Asymptotic distribution theory for parameter estimates is developed and procedures for making statistical inference are evaluated with simulation studies. We illustrate our methods with data from the Cardiovascular Health Study, relating the FR-score measured at enrollment to subsequent risk of cardiovascular events.

Khalili et al. (1) evaluated the clinical utility of the Framingham cardiovascular risk model in a Middle Eastern cohort study. We commend the authors for their use of highly relevant measures of model performance, the net benefit (2, 3) and net benefit fraction (4, 5). However, we would like to make several points regarding their interpretation and reporting of these measures. Interpretation of the net benefit and net benefit fraction The population net benefit (NB) of a risk model, used to classify individuals as being at high 10-year risk of cardiovascular disease using the risk threshold rH, is (1) where B is the benefit of classifying as high-risk a subject who would develop cardiovascular disease absent treatment (a “cardiovascular event”), H is the harm of classifying as high-risk a subject who would not develop cardiovascular disease absent treatment (a “cardiovascular nonevent”), TPR(rH) and FPR(rH) are the true- and false-positive rates associated with this risk threshold, and ρ is the proportion of subjects who experience cardiovascular events. Vickers and Elkin (2) set B to 1, so that the units are true positive classifications, and use the optimal risk threshold rH = H/H + B (6) in order to rewrite equation 1 as (2) Baker et al. (4, 5) scale equation 2 by ρ, which is the maximum possible net benefit for a model that classifies as high-risk all cardiovascular events and all others as low-risk. The net benefit fraction (NBF) is therefore (3) Khalili et al. interpret NBF(0.2) = 0.156 for the Framingham risk model applied to women as “the fraction of the incidence rate that could be predicted and prevented” (1, p. 178). However, this is not a valid interpretation. Their interpretation seems to relate to the TPR of the model, whereas NBF(0.2) = 0.156 is more accurately interpreted as a discounted TPR; the TPR is offset by the FPR where false-positives are put on the same scale as true-positives (7). We prefer to interpret NBF(0.2) = 0.156 as the same benefit that would be achieved by classifying as high-risk 15.6% of cardiovascular events and none of the nonevents (2, 3). The authors' interpretation also suggests that all cardiovascular events classified as high-risk would necessarily be prevented; however, this would depend on the efficacy of subsequent interventions. Net benefit for different subsets of the population Net benefit curves, which plot the net benefit in equation 2 versus the risk threshold rH (2, 3), are also used by Khalili et al. The authors recommend these curves for examining the net benefit for a range of risk thresholds, given that different patient subgroups may have different side effects of interventions and therefore different risk thresholds. However, the net benefit curve shows the total population net benefit at each risk threshold. Two points on the curve cannot be interpreted as the net benefits for 2 different subpopulations because, if the risk distribution in the subpopulation differs from the risk distribution in the whole population, the net benefit in the subpopulation is different from the net benefit in the whole population at that threshold. For example, men and women in the Middle Eastern cohort have different values for TPR, FPR, and ρ and therefore have net benefits that differ from the population net benefit shown on the net benefit curve. Recommendation We have several practical recommendations. The first is to stratify any analysis by important patient characteristics that affect risk threshold—for example, gender as in the study by Khalili et al. (1). The second is to report the net benefit fraction together with its constituents, the TPR and FPR, over a range of risk thresholds. Seeing these individual components helps in digesting the net benefit. Knowing the TPR and FPR is also helpful when risk thresholds are chosen “irrationally.” That is, the net benefit in equation 2 assumes that the risk threshold reflects the cost-benefit ratio, rH = H/H + B. However, individuals or policy-makers may choose risk thresholds in ways that do not reflect the cost-benefit ratio. For example, in their commentary accompanying Khalili et al.'s article, D'Agostino and Pencina (8) suggest that the risk threshold rH = 0.2 was chosen without regard to cost-benefit considerations. Reporting the TPR and FPR as a function of risk threshold allows the reader to choose a threshold and to perform any calculus desired to take into consideration both components of model performance.

Selecting the best treatment for a patient's disease may be facilitated by evaluating clinical characteristics or biomarker measurements at diagnosis. We consider how to evaluate the potential impact of such measurements on treatment selection algorithms. For example, magnetic resonance neurographic imaging is potentially useful for deciding whether a patient should be treated surgically for Carpal Tunnel Syndrome or should receive less-invasive conservative therapy. We propose a graphical display, the selection impact (SI) curve that shows the population response rate as a function of treatment selection criteria based on the marker. The curve can be useful for choosing a treatment policy that incorporates information on the patient's marker value exceeding a threshold. The SI curve can be estimated using data from a comparative randomized trial conducted in the population as long as treatment assignment in the trial is independent of the predictive marker. Estimating the SI curve is therefore part of a post hoc analysis to determine whether the marker identifies patients that are more likely to benefit from one treatment over another. Nonparametric and parametric estimates of the SI curve are proposed in this article. Asymptotic distribution theory is used to evaluate the relative efficiencies of the estimators. Simulation studies show that inference is straightforward with realistic sample sizes. We illustrate the SI curve and statistical inference for it with data motivated by an ongoing trial of surgery versus conservative therapy for Carpal Tunnel Syndrome.

The idea of using measurements such as biomarkers, clinical data, or molecular biology assays for classification and prediction is popular in modern medicine. The scientific evaluation of such measures includes assessing the accuracy with which they predict the outcome of interest. Receiver operating characteristic curves are commonly used for evaluating the accuracy of diagnostic tests. They can be applied more broadly, indeed to any problem involving classification to two states or populations (D= 0 or 1). We show that the ROC curve can be interpreted as a cumulative distribution function for the discriminatory measure Y in the affected population (D= 1) after Y has been standardized to the distribution in the reference population (D= 0). The standardized values are called placement values. If the placement values have a uniform(0, 1) distribution, then Y is not discriminatory, because its distribution in the affected population is the same as that in the reference population. The degree to which the distribution of the standardized measure differs from uniform(0, 1) is a natural way to characterize the discriminatory capacity of Y and provides a nontraditional interpretation for the ROC curve. Statistical methods for making inference about distribution functions therefore motivate new approaches to making inference about ROC curves. We demonstrate this by considering the ROC-GLM regression model and observing that it is equivalent to a regression model for the distribution of placement values. The likelihood of the placement values provides a new approach to ROC parameter estimation that appears to be more efficient than previously proposed methods. The method is applied to evaluate a pulmonary function measure in cystic fibrosis patients as a predictor of future occurrence of severe acute pulmonary infection requiring hospitalization. Finally, we note the relationship between regression models for the mean placement value and recently proposed models for the area under the ROC curve which is the classic summary index of discrimination.

Plasma ghrelin levels rise before meals and fall rapidly afterward. If ghrelin is a physiological meal-initiation signal, then a large oral caloric load should suppress ghrelin levels more than a small caloric load, and the request for a subsequent meal should be predicted by recovery of the plasma ghrelin level. To test this hypothesis, 10 volunteers were given, at three separate sessions, liquid meals (preloads) with widely varied caloric content (7.5%, 16%, or 33% of total daily energy expenditure) but equivalent volume. Preloads were consumed at 0900 h, and blood was sampled every 20 min from 0800 h until 80 min after subjects spontaneously requested a meal. The mean (+/- SE) intervals between ingestion of the 7.5%, 16%, and 33% preloads and the subsequent voluntary meal requests were 247 +/- 24, 286 +/- 20, and 321 +/- 27 min, respectively (P = 0.015), and the nadir plasma ghrelin levels were 80.2 +/- 2.8%, 72.7 +/- 2.7%, and 60.8 +/- 2.7% of baseline (the 0900 h value), respectively (P0.001). A Cox regression analysis failed to show a relationship between ghrelin profile and the spontaneous meal request. We conclude that the depth of postprandial ghrelin suppression is proportional to ingested caloric load but that recovery of plasma ghrelin is not a critical determinant of intermeal interval.

SUMMARY The positive and negative predictive values are standard ways of quantifying predictive accuracy when both the outcome and the prognostic factor are binary. Methods for comparing the predictive values of two or more binary factors have been discussed previously (Leisenring et al., 2000, Biometrics 56, 345–351). We propose extending the standard definitions of the predictive values to accommodate prognostic factors that are measured on a continuous scale and suggest a corresponding graphical method to summarize predictive accuracy. Drawing on the work of Leisenring et al. we make use of a marginal regression framework and discuss methods for estimating these predictive value functions and their differences within this framework. The methods presented in this paper have the potential to be useful in a number of areas including the design of clinical trials and health policy analysis.

Accurate diagnosis of disease is a critical part of health care. New diagnostic and screening tests must be evaluated based on their abilities to discriminate diseased from nondiseased states. The partial area under the receiver operating characteristic (ROC) curve is a measure of diagnostic test accuracy. We present an interpretation of the partial area under the curve (AUC), which gives rise to a nonparametric estimator. This estimator is more robust than existing estimators, which make parametric assumptions. We show that the robustness is gained with only a moderate loss in efficiency. We describe a regression modeling framework for making inference about covariate effects on the partial AUC. Such models can refine knowledge about test accuracy. Model parameters can be estimated using binary regression methods. We use the regression framework to compare two prostate-specific antigen biomarkers and to evaluate the dependence of biomarker accuracy on the time prior to clinical diagnosis of prostate cancer.

SUMMARY In early detection of disease, combinations of biomarkers promise improved discrimination over diagnostic tests based on single markers. An example of this is in prostate cancer screening, where additional markers have been sought to improve the specificity of the conventional Prostate-Specific Antigen (PSA) test. A marker of particular interest is the percent free PSA. Studies evaluating the benefits of percent free PSA reflect the need for a methodological approach that is statistically valid and useful in the clinical setting. This article presents methods that address this need. We focus on and-or combinations of biomarker results that we call logic rules and present novel definitions for the ROC curve and the area under the curve (AUC) that are applicable to this class of combination tests. Our estimates of the ROC and AUC are amenable to statistical inference including comparisons of tests and regression analysis. The methods are applied to data on free and total PSA levels among prostate cancer cases and matched controls enrolled in the Physicians’ Health Study.

When multiple diagnostic tests are performed on an individual or multiple disease markers are available it may be possible to combine the information to diagnose disease. We consider how to choose linear combinations of markers in order to optimize diagnostic accuracy. The accuracy index to be maximized is the area or partial area under the receiver operating characteristic (ROC) curve. We propose a distribution-free rank-based approach for optimizing the area under the ROC curve and compare it with logistic regression and with classic linear discriminant analysis (LDA). It has been shown that the latter method optimizes the area under the ROC curve when test results have a multivariate normal distribution for diseased and non-diseased populations. Simulation studies suggest that the proposed non-parametric method is efficient when data are multivariate normal.The distribution-free method is generalized to a smooth distribution-free approach to: (i) accommodate some reasonable smoothness assumptions; (ii) incorporate covariate effects; and (iii) yield optimized partial areas under the ROC curve. This latter feature is particularly important since it allows one to focus on a region of the ROC curve which is of most relevance to clinical practice. Neither logistic regression nor LDA necessarily maximize partial areas. The approaches are illustrated on two cancer datasets, one involving serum antigen markers for pancreatic cancer and the other involving longitudinal prostate specific antigen data.

Disease screening is a fundamental part of health care. To evaluate the accuracy of a new screening modality, ideally the results of the screening test are compared with those of a definitive diagnostic test in a set of study subjects. However, definitive diagnostic tests are often invasive and cannot be applied to subjects whose screening tests are negative for disease. For example, in cancer screening, the assessment of true disease status requires a biopsy sample, which for ethical reasons can only be obtained if a subject's screening test indicates presence of cancer. Although the absolute accuracy of screening tests cannot be evaluated in such circumstances, it is possible to compare the accuracies of screening tests. Specifically, using relative true positive rate (the ratio of the true positive rate of one test to another) and relative false positive rate (the ratio of the false positive rates of two tests) as measures of relative accuracy, we show that inference about relative accuracy can be made from such studies. Analogies with case-control studies can be drawn where inference about absolute risk cannot be made, but inference about relative risk can. In this paper, we develop a marginal regression analysis framework for making inference about relative accuracy when only screen positives are followed for true disease. In this context factors influencing the relative accuracies of tests can be evaluated. It is important to determine such factors in order to understand circumstances in which one test is preferable to another. The methods are applied to two cancer screening studies, one concerning the effect of race on screening for prostate cancer and the other concerning the effect of tumour grade on the detection of cervical cancer with cytology versus cervicography screening.

SUMMARY Disease prevalence is ideally estimated using a ‘gold standard’ to ascertain true disease status on all subjects in a population of interest. In practice, however, the gold standard may be too costly or invasive to be applied to all subjects, in which case a two-phase design is often employed. Phase 1 data consisting of inexpensive and non-invasive screening tests on all study subjects are used to determine the subjects that receive the gold standard in the second phase. Naive estimates of prevalence in two-phase studies can be biased (verification bias). Imputation and re-weighting estimators are often used to avoid this bias. We contrast the forms and attributes of the various prevalence estimators. Distribution theory and simulation studies are used to investigate their bias and efficiency. We conclude that the semiparametric efficient approach is the preferred method for prevalence estimation in two-phase studies. It is more robust and comparable in its efficiency to imputation and other re-weighting estimators. It is also easy to implement. We use this approach to examine the prevalence of depression in adolescents with data from the Great Smoky Mountain Study.

With recent advances in mass spectrometry techniques, it is now possible to investigate proteins over a wide range of molecular weights in small biological specimens. This advance has generated data-analytic challenges in proteomics, similar to those created by microarray technologies in genetics, namely, discovery of 'signature' protein profiles specific to each pathologic state (e.g. normal vs. cancer) or differential profiles between experimental conditions (e.g. treated by a drug of interest vs. untreated) from high-dimensional data. We propose a data-analytic strategy for discovering protein biomarkers based on such high-dimensional mass spectrometry data. A real biomarker-discovery project on prostate cancer is taken as a concrete example throughout the paper: the project aims to identify proteins in serum that distinguish cancer, benign hyperplasia, and normal states of prostate using the Surface Enhanced Laser Desorption/Ionization (SELDI) technology, a recently developed mass spectrometry technique. Our data-analytic strategy takes properties of the SELDI mass spectrometer into account: the SELDI output of a specimen contains about 48,000 (x, y) points where x is the protein mass divided by the number of charges introduced by ionization and y is the protein intensity of the corresponding mass per charge value, x, in that specimen. Given high coefficients of variation and other characteristics of protein intensity measures (y values), we reduce the measures of protein intensities to a set of binary variables that indicate peaks in the y-axis direction in the nearest neighborhoods of each mass per charge point in the x-axis direction. We then account for a shifting (measurement error) problem of the x-axis in SELDI output. After this pre-analysis processing of data, we combine the binary predictors to generate classification rules for cancer, benign hyperplasia, and normal states of prostate. Our approach is to apply the boosting algorithm to select binary predictors and construct a summary classifier. We empirically evaluate sensitivity and specificity of the resulting summary classifiers with a test dataset that is independent from the training dataset used to construct the summary classifiers. The proposed method performed nearly perfectly in distinguishing cancer and benign hyperplasia from normal. In the classification of cancer vs. benign hyperplasia, however, an appreciable proportion of the benign specimens were classified incorrectly as cancer. We discuss practical issues associated with our proposed approach to the analysis of SELDI output and its application in cancer biomarker discovery.

Staging of patients with lung cancer to determine operability is intended to efficiently limit futile thoracotomies without denying possibly curative surgery. Currently available staging tests are imperfect alone and in combination. Imaged suspected metastases often require tissue confirmation before surgery can be denied. Endoscopic ultrasound (EUS) may help identify inoperable patients by providing tissue proof of inoperability in a single staging test, with similar sensitivity for identifying inoperable patients as other staging tests. Therefore, we compared computed tomography, positron emission tomography (PET), and EUS with fine-needle aspiration under conscious sedation, each test interpreted blinded with respect to the other tests, for identifying inoperable patients in a consecutive cohort of 79 potentially operable patients with suspected or proven lung cancer. An economic analysis was also performed. Thirty-nine patients were found inoperable (a 40th patient's inoperability was missed by all preoperative staging tests). The sensitivity of computerized tomography was 43%. PET and EUS each had similar sensitivities (68 and 63%, respectively) and similar negative predictive values (64 and 68%, respectively), but EUS's superior specificity (100 vs. 72% for PET) and considerably lower expense means it may be preferred to PET early in staging to identify inoperable patients.